M.Phil. ProjectYear 2007
Title: A Search Methodology for Near-Optimal Artificial Neural Networks (Uma Metodologia de Busca por Redes Neurais Artificiais Quase-Ótimas.).
Supervisor: Prof. Teresa B. Ludermir.
Keywords: artificial neural networks, genetic algorithms, memetic algorithms, near-optimal nets, on-line parameter optimization, numeric parameter optimization.
Funding: Brazilian Council for Scientific and Technological Development (CNPq).
This works introduces a methodology for automatically searching near-optimal Artificial Neural Networks (ANN) for classification problems. The aim is to search networks with a simple architecture that learn faster and with good classification capabilities, in other words, near-optimal networks. The motivation for development of the current work is centered on the difficulties of searching near-optimal ANNs using manual methods. These difficulties are due to the large amount of neural network parameters that must be adjusted in order to produce a good correlation between parameters that contribute toward the obtainment of simple networks with high performance.
The automatic search of near-optimal networks includes information such as initial weights, hidden layers, number of nodes per layer, activation functions and learning algorithms of fully connected Multi-Layer Perceptron (MLP). The search mechanism is made up of a combination of Genetic Algorithms (GA) and ANNs, whereby a global search is first executed with AGs using ANN parameters and a local search is then executed using ANN learning algorithms in order to refine and evaluate the solution encountered. This kind of search has been established and has obtained great results in previous work found in the literature. The difference in the current method is the focus on simplified architectures, with a high classification performance in few training epochs.
Experiments are performed are carried out with the proposed method, using five well-known classifications problems: Cancer, Glass, Heart, Horse and Diabetes. The results demonstrate the greater effectiveness of the method in searching near-optimal ANNs in comparison to the manual method and other methods in the literature. The networks found for each domain problem exhibit low complexity as well as low classification error. These results are extremely important in demonstrating the capacity of the developed method and justify efforts in the development of methods for searching near-optimal ANNs.